first-principles molecular dynamics - uniudgiannozz/slides/venice15.pdffirst-principles molecular...
TRANSCRIPT
First-principles molecular dynamics
P. GiannozziDept of Chemistry, Physics, Environment, University of Udine, Italy
XIX School of Pure and Applied Biophysics, Venice, 26 January 2015
Outline
1. First-principles molecular dynamics: when and what?
2. A quick look at the theoretical basis, with emphasis on Density-Functional Theory
3. Born-Oppenheimer and Car-Parrinello Molecular Dynamics
4. Technicalites you should be aware on: plane waves, pseudopotentials
5. Software and needed hardware
– Typeset by FoilTEX –
1. What is First-Principles Molecular Dynamics?
In First-Principles Molecular Dynamics (FPMD) the forces acting on nuclei arecomputed from the electronic states, i.e. by solving, in principle exactly, themany-body Schrodinger equation for electrons (nuclei are approximated by classicalparticles).
FPMD is the method of choice whenever “force fields” do not give a sufficientlyaccurate or reliable description of the system under study, or of a specificphenomenon: typically, whenever bonds are broken or formed, or in presence ofcomplex bonding, e.g.: transition metals.
With FPMD, all techniques used in classical (with force fields) MD can be used;moreover, one has access to the electronic structure and charge density, can inprinciple compute excitation spectra: however, the heavy computational load ofFPMD, in spite of available efficient implementations, limits its range of applicationto systems containing O(1000) atoms max, for rather short times (tens of ps).
An example of application
Below, the simplest model for themyoglobin active site: iron-porphyrin-imidazole complex. Yellow: Fe. Dark Gray:C. Blue: N. Light gray: H.
To the right: extended model of themyoglobin active site (including the 13residues in a 8A sphere centered on Fe,terminated with NH2 groups, containing332 atoms, 902 electrons). Red: O, allother atoms as in the above picture.
The need for first-principles MD
• High-level theoretical description needed, dueto the presence of transition metal atoms
• Complex energy landscape: many possible spinstates and local minima, not easy to guess fromstatic (single-point) calculations – MolecularDynamics may be better suited.
Interest: respiration, photosynthesis, enzymaticcatalysis...
Note however that sizable systems (hundreds of atoms in this case) areneeded in order to accurately describe the effect of surrounding environment:efficient first-principles techniques are needed (P. Giannozzi, F. de Angelis, andR. Car, J. Chem. Phys. 120, 5903 (2004); D. Marx and J. Hutter,Modern Methods and Algorithms of Quantum Chemistry, John von NeumannInstitute for Computing, Julich, NIC Series, Vol.3, pp. 329-477 (2000)http://juser.fz-juelich.de/record/152459/files/FZJ-2014-02061.pdf)
2. The basics: Born-Oppenheimer approximation
The behavior of a system of interacting electrons r and nuclei R is determined bythe solutions of the time-dependent Schrodinger equation:
ih∂Φ(r,R; t)
∂t=
−∑µ
h2
2Mµ
∂2
∂R2µ
−∑i
h2
2m
∂2
∂r2i
+ V (r,R)
Φ(r,R; t)
where V (r,R) is the potential describing the coulombian interactions:
V (r,R) =∑µ>ν
ZµZνe2
|Rµ −Rν|−∑i,µ
Zµe2
|ri −Rµ|+∑i>j
e2
|ri − rj|
≡ Vnn(R) + Vne(r,R) + Vee(r)
Born-Oppenheimer (or adiabatic) approximation (valid for Mµ >> m):
Φ(r,R; t) ' Φ(R)Ψ(r|R)e−iEt/h
NB: in order to keep the notation simple: r ≡ (r1, .., rN), R ≡ (R1, ..,Rn); spinand relativistic effects are not included
Potential Energy Surface
The Born-Oppenheimer approximation allows to split the original problem into anelectronic problem depending upon nuclear positions:(
−∑i
h2
2m
∂2
∂r2i
+ V (r,R)
)Ψ(r|R) = E(R)Ψ(r|R)
and a nuclear problem under an effective interatomic potential, determined by theelectrons: −∑
µ
h2
2Mµ
∂2
∂R2µ
+ E(R)
Φ(R) = EΦ(R)
E(R) defines the Potential Energy Surface (PES). The PES determines the motionof atomic nuclei and the equilibrium geometry: the forces Fµ on nuclei, defined as
Fµ = −∂E(R)
∂Rµ,
vanish at equilibrium: Fµ = 0.
Solving the electronic problem: Hartree-Fock
Let us write the many-body electronic wave-function as a Slater determinant:
Ψ(r1, r2, ...rN) =1√N !
∣∣∣∣∣∣∣∣ψ1(1) . . . ψ1(N). . .. . .
ψN(1) . . . ψN(N)
∣∣∣∣∣∣∣∣for N electrons, where (i) ≡ (ri, σi), σi is the spin of the i−th electron, the ψistates are all different. Minimization of E yields the Hartree-Fock equations.
For the “restricted” case (pairs of states with opposite spin):
− h2
2m∇2ψi(r) + V (r)ψi(r) + VH(r)ψi(r) + (Vxψi)(r) = εiψi(r)
where i = 1, ..., N/2, V (r) is the external (nuclear) potential acting on each electron:
V (r) = −∑µ
Zµe2
|r−Rµ|,
VH is the Hartree potential:
VH(r) = 2∑j
∫|ψj(r′)|2
e2
|r− r′|dr′
Vx is the (nonlocal) exchange potential:
(Vxψi)(r) = −∑j
∫ψj(r)
e2
|r− r′|ψ∗j (r
′)ψi(r′)dr′.
These are integro-differential equations that must be solved self-consistently.
In practice, HF is not a very good approximation. More accurate resultsrequire perturbative corrections (Møller-Plesset) or to add more Slater determinants(Configuration Interaction) in order to find more of the correlation energy, definedas “the difference between the true and the Hartree-Fock energy”.
Solving HF and post-HF equations is the “traditional” approach of QuantumChemistry. An alternative approach, based on the charge density, is provided byDensity-Functional Theory (DFT)
Solving the electronic problem: Hohenberg-Kohn theorem
Let us introduce the ground-state charge density n(r). For N electrons:
n(r) = N
∫|Ψ(r, r2, ...rN)|2dr2...drN .
The Hohenberg-Kohn theorem (1964) can be demonstrated: there is a uniquepotential V (r,R) having n(r) as ground-state charge density. Consequences:
• The energy can be written as a functional of n(r):
E[n(r)] = F [n(r)] +
∫n(r)V (r)dr
where F [n(r)] is a universal functional of the density, V (r) is the external(nuclear) potential acting on each electron:
V (r) = −∑µ
Zµe2
|r−Rµ|.
• E[n(r)] is minimized by the ground-state charge density n(r).
Note: the energy E defined above does not include nuclear-nuclear interaction terms
Density-Functional Theory: Kohn-Sham approach
Let us introduce the orbitals ψi for an auxiliary set of non-interacting electronswhose charge density is the same as that of the true system:
n(r) =∑i
|ψi(r)|2, 〈ψi|ψj〉 = δij
Let us rewrite the energy functional in a more manageable way as:
E = Ts[n(r)] + EH[n(r)] + Exc[n(r)] +
∫n(r)V (r)dr
where Ts[n(r)] is the kinetic energy of the non-interacting electrons:
Ts[n(r)] = − h2
2m
∑i
∫ψ∗i (r)∇2ψi(r)dr,
EH[n(r)] is the Hartree energy, due to electrostatic interactions:
EH[n(r)] =e2
2
∫n(r)n(r′)
|r− r′|drdr′,
Exc[n(r)] is called exchange-correlation energy (a reminiscence from the Hartree-Fock theory) and includes all the remaining (unknown!) energy terms.
Minimization of the energy with respect to ψi yields the Kohn-Sham (KS) equations:(− h
2
2m∇2 + V (r) + VH(r) + Vxc(r)
)︸ ︷︷ ︸
HKS
ψi(r) = εiψi(r),
where the Hartree and exchange-correlation potentials:
VH(r) =δEH[n(r)]
δn(r)= e2
∫n(r′)
|r− r′|dr′, Vxc(r) =
δExc[n(r)]
δn(r)
depend self-consistently upon the ψi via the charge density.
The energy can be rewritten in an alternative form using the KS eigenvalues εi:
E =∑i
εi − EH[n(r)]−∫n(r)Vxc(r)dr + Exc[n(r)]
Exchange-correlation functionals: simple approximations
What is Exc[n(r)]? Viable approximations needed to turn DFT into a useful tool.The first, ”historical” approach (1965) is the Local Density Approximation (LDA):replace the energy functional with a function of the local density n(r),
Exc =
∫n(r)εxc(n(r))dr, Vxc(r) = εxc(n(r)) + n(r)
dεxc(n)
dn
∣∣∣∣n=n(r)
where εxc(n) is calculated for the homogeneous electron gas of uniform density n(e.g. using Quantum Monte Carlo) and parameterized.
Generalized Gradient Approximation (GGA): A more recent class of functionalsdepending upon the local density and its local gradient |∇n(r)|,
Exc =
∫n(r)εGGA (n(r), |∇n(r)|) dr
There are many flavors of GGA, yielding similar (but slightly different) results. GGAis the ”standard” functional in most FPMD calculations, with excellent price-to-performance ratio, but some noticeable shortcomings.
Spin-polarized extension: LSDA
Simplest case: assume a unique quantization axis for spin. Energy functional:
E ≡ E[n+(r), n−(r)] = Ts +
∫n(r)V (r)dr + EH + Exc[n+(r), n−(r)]
nσ(r) = charge density with spin polarization σn(r) = n+(r) + n−(r) total charge density.
Minimization of the above functional yields the Kohn-Sham equations:[− h
2
2m∇2 + V (r) + e2
∫n(r′)
| r− r′ |dr′ + V σxc(r)
]ψσi (r) = εσi ψ
σi (r)
Exchange-correlation potential and charge density:
V σxc(r) =δExcδnσ(r)
, nσ(r) =∑i
fσi |ψσi (r)|2
Note the extension to fractional occupancies (i.e. metallic systems): 0 ≤ fσi ≤ 1.Noncolinear magnetism (no fixed axis for magnetization) can also be described.
“Standard” DFT: advantages and shortcomings
+ Computationally convenient: calculations in relatively complex condensed-matter systems become affordable (GGA marginally more expensive than LDA)
+ Excellent results in terms of prediction of atomic structures, bond lengths, latticeparameters (within 1÷ 2%), binding and cohesive energies (5 to 10% GGA; LDAmuch worse, strongly overestimates), vibrational properties. Especially good forsp−bonded materials, may work well also in more ”difficult” materials, such astransition metal compounds
– The infamous band gap problem: εc − εv (or HOMO-LUMO in quantumchemistry parlance) wildly underestimates the true band gap (mostly due tomissing discontinuity in Vxc for integer number of electrons)
– Serious trouble in dealing with strongly correlated materials, such as e.g.magnetic materials (trouble mostly arising from spurious self-interaction)
– No van der Waals interactions in any functional based on the local density andgradients: van der Waals is nonlocal, cannot depend upon charge overlap
Advanced DFT functionals
• DFT+U. LDA and GGA are often unable to find the correct occupancy of atomic-like electronic states, leading to qualitatively wrong results – but when occupancyis correct, results are not bad. DFT+U adds a Hubbard-like term, accounting forstrong Coulomb correlations in systems with highly localized, atomic-like states:
EDFT+U [n(r)] = EDFT [n(r)] + EU [n(r)],
where (simplified, rotationally invariant form):
EU [n(r)] =U
2
∑σ
Tr[nσ(1− nσ)],
U being a (system-dependent) Coulomb repulsion (typically a few eV) and nσ isthe matrix of orbital occupancies for a set of atomic-like states φm:
nσmm′ =∑σ
∑i
fσi 〈ψσi |Pmm′|ψσi 〉, Pmm′ = |φm〉〈φm′|
DFT+U is a “quick-and-dirty” but effectivesolution for a deep problem of DFT: the lackof discontinuity in approximated functionals, dueto incomplete self-interactions cancellation, favorsfractionary occupancy. DFT+U also improvesthe gap and level alignement in heterostructures,at the price of introducing an adjustable parameter U
figure: Cococcioni&de Gironcoli, PRB 71, 035105 (2005)
• Meta-GGA. Has a further dependence upon the non-interacting kinetic energydensity τs(r):
Exc =
∫εmGGA (n(r), |∇n(r)|, τs(r)) dr
where
τs(r) =1
2
∑i
|∇2ψi(r)|2
Meta-GGA may yield very good results and is not in principle much more expensivethan GGA but it is numerically very nasty (at least in my experience)
• Hybrid functionals, such as B3LYP or PBE0, containing some amount of exactexchange, as in Hartree-Fock theory (spin-restricted):
E = Ts +
∫n(r)V (r)dr + EH −
e2
2
∑i,j
∫ψ∗i (r)ψj(r)ψ∗j (r
′)ψi(r′)
|r− r′|drdr′︸ ︷︷ ︸
EHFx
In Hybrid DFT:
E = Ts +
∫n(r)V (r)dr + EH +
Ehybxc︷ ︸︸ ︷αxE
HFx + (1− αx)EDFTx + EDFTc ,
with αx = 20 ÷ 30%. This is the method of choice in Quantum Chemistry,yielding very accurate results and correcting most GGA errors.
In FPMD, however, hybrid functionals are computationally very heavy (due tousage of plane waves). Moreover, an adjustable parameter is introduced.
• Nonlocal vdW functionals, accounting for van der Waals (dispersive) forces,with a reasonable computational overhead.
Towards electronic ground state (fixed nuclei)
Possible methods to find the DFT ground state:
1. By the self-consistent solution of the Kohn-Sham equations
HKSψi ≡ (T + V + VH[n] + Vxc[n])ψi = εiψi
where
– n(r) =∑i
fi|ψi(r)|2 is the charge density, fi are occupation numbers
– V is the nuclear (pseudo-)potential acting on electrons (may be nonlocal)
– VH[n] is the Hartree potential, VH(r) = e2
∫n(r′)
|r− r′|dr′
– Vxc[n] is the exchange-correlation potential. For the simplest case, LDA, Vxc[n]is a function of the charge density at point r: Vxc(r) ≡ µxc(n(r))
Orthonormality constraints 〈ψi|ψj〉 = δij automatically hold.
2. By constrained global minimization of the energy functional
E[ψ] =∑i
fi〈ψi|T + V |ψi〉+ EH[n] + Exc[n]
under orthonormality constraints 〈ψi|ψj〉 = δij, i.e. minimize:
E[ψ,Λ] = E[ψ]−∑ij
Λij (〈ψi|ψj〉 − δij)
where
– V , n(r) are defined as before, ψ ≡ all occupied Kohn-Sham orbitals– Λij are Lagrange multipliers, Λ ≡ all of them
– EH[n] is the Hartree energy, EH =e2
2
∫n(r)n(r′)
|r− r′|drdr′
– Exc[n] is the exchange-correlation energy. For the simplest case, LDA,
Exc =
∫n(r)εxc(n(r))dr where εxc is a function of n(r).
Towards electronic ground state II
In the global-minimization approach, we need to compute the gradients of the energyfunctional, that is, HKSψ products:
δE[ψ,Λ]
δψ∗i= HKSψi −
∑j
Λijψj
In the self-consistent approach with iterative diagonalization, the basic ingredient isthe same: HKSψ products.
We further need to compute
1. the charge density from the ψi’s, and
2. the potential from the charge density.
Typically, constrained global minimization is the preferred technique for FPMD.
Calculation of the total energy
Once the electronic ground state is reached, the total energy of the system can becalculated:
E =∑i
fi〈ψi|T + V |ψi〉+ EH[n] + Exc[n] + Enn
where Enn is the repulsive contribution from nuclei to the energy:
Enn =e2
2
∑µ6=ν
ZµZν| Rµ −Rν |
Equivalent expression for the energy, using Kohn-Sham eigenvalues:
E =∑i
fiεi − EH[n] + Exc[n(r)]−∫n(r)Vxc[n(r)]dr + Enn
The total energy depends upon all atomic positions Rµ (the PES is nothing but thetotal energy for all Rµ).
Hellmann-Feynman Forces
Forces on atoms are the derivatives of the total energy wrt atomic positions. TheHellmann-Feynman theorem tells us that forces are the expectation value of thederivative of the external potential only:
Fµ = − ∂E
∂Rµ= −
∑i
fi〈ψi|∂V
∂Rµ|ψi〉 = −
∫n(r)
∂V
∂Rµdr
the rightmost expression being valid only for local potentials, V ≡ V (r) (the one atthe left is more general, being valid also for nonlocal potentials V ≡ V (r, r′)).
Demonstration (simplified). In addition to the explicit derivative of the external potential (first term),
there is an implicit dependency via the derivative of the charge density:
∂E
∂Rµ
=
∫n(r)
∂V
∂Rµ
dr +
∫δE
δn(r)
∂n(r)
∂Rµ
dr
The green term cancels due to the variational character of DFT: δE/δn(r) = µ, constant.
The calculation of the Hellmann-Feynman forces is straightforward (in principle, notnecessarily in practice!) once the ground-state electronic structure is available.
3.1 Born-Oppenheimer Molecular Dynamics
Let us assume classical behavior for the nuclei and electrons in the ground state.We introduce a classical Lagrangian:
L =1
2
∑µ
MµR2µ − E(R)
describing the motion of nuclei. The equations of motion:
d
dt
∂L
∂Rµ
− ∂L
∂Rµ= 0, Pµ =
∂L
∂Rµ
are nothing but usual Newton’s equations:
Pµ ≡MµVµ, MµVµ = Fµ,
that can be discretized and solved by integration.This procedure defines Molecular Dynamics “onthe Born-Oppenheimer surface”, with electronsalways at their instantaneous ground state.
Discretization of the equation of motion
Like in classical MD, the equation of motions can be discretized using the Verletalgorithm:
Rµ(t+ δt) = 2Rµ(t)−Rµ(t− δt) +δt2
MµFµ(t) +O(δt4)
Vµ(t) =1
2δt[Rµ(t+ δt)−Rµ(t− δt)] +O(δt3).
or the Velocity Verlet:
Vµ(t+ δt) = Vµ(t) +δt
2Mµ[Fµ(t) + Fµ(t+ δt)]
Rµ(t+ δt) = Rµ(t) + δtVµ(t) +δt2
2MµFµ(t).
Both sample the microcanonical ensemble, or NVE: the energy (mechanical energy:kinetic + potential) is conserved.
Thermodynamical averages
Averages (ρ = probability of a microscopic state):
〈A〉 =
∫ρ(R,P)A(R,P)dRdP
are usually well approximated by averages over time:
limT→∞
AT → 〈A〉
and the latter by discrete average over a trajectory:
AT =1
T
∫ T
0
A(R(t),P(t))dt ' 1
M
M∑n=1
A(tn), tn = nδt, tM = Mδt = T.
Costant-Temperature and constant-pressure dynamics
• Molecular Dynamics with Nose Thermostats samples the canonical ensemble(NVT): the average temperature
〈N∑µ=1
P2µ
2Mµ〉NV T =
3
2NkBT
is fixed (the instantaneous value has wide oscillations around the desired value)
• Molecular Dynamics with variable cell samples the NPT ensemble: the averagepressure is fixed
In both cases, additional (fictitious) degrees of freedom are added to the system
Technicalities
• time step as big as possible, but small enough to follow nuclear motion with littleloss of accuracy. Rule of thumb: δt ∼ 0.01−0.1δtmax, where δtmax = 1/ωmax =period of the fastest phonon (vibrational) mode.
• calculations of forces must be very well converged (good self-consistency needed)at each time step or else a systematic drift of the conserved energy will appear
Note that:
– the error on DFT energy is a quadratic function of the self-consistency error ofthe charge density (because energy has a minimum in correspondence to theself-consistent charge)
– the error for DFT forces is a linear function of the self-consistency error of thecharge density
As a consequence, Born-Oppenheimer MD is usually computationally heavy
3.2 Car-Parrinello Molecular Dynamics
The idea: introduce a fictitious electron dynamics that keeps the electrons close tothe ground state. The electron dynamics is faster than the nuclear dynamics andaverages out the error, but not too fast so that a reasonable time step can be used
Car-Parrinello Lagrangian:
L =m∗
2
∑i
∫|ψi(r)|
2dr+
1
2
∑µ
MµR2µ−E[R, ψ]+
∑i,j
Λij
(∫ψ∗i (r)ψj(r)dr− δij
)generates equations of motion:
m∗ψi = Hψi−∑j
Λijψj, MµRµ = Fµ ≡ −∂E
∂Rµ
m∗ = fictitious electronic massΛij = Lagrange multipliers, enforcingorthonormality constraints.Very effective, but requires a judicious choice ofsimulation parameters.
Car-Parrinello Molecular Dynamics (2)
• electronic degrees of freedom ψi are the expansion coefficients of KS orbitals intoa suitable basis set (typically Plane Waves for technical reasons)
• ”forces” on electrons are determined by the KS Hamiltonian calculated fromcurrent values of ψi and of Rµ
• ”forces” acting on nuclei have the Hellmann-Feynman form:
∂E
∂Rµ=∑i
〈ψi|∂V
∂Rµ|ψi〉
but they slightly differ from ”true” forces (ψi are not exact ground-state orbitals)
• The simulation is performed using classical MD technology (e.g. Verlet) on bothnuclear positions and electronic degrees of freedom (Kohn-Sham orbitals)
• Orthonormality constraints are imposed exactly at each time step, using aniterative procedure
Car-Parrinello technicalities
• Starting point: bring the electrons to the ground state at fixed nuclear positions– this can be achieved using damped dynamics.
• Next step is often to bring the system to an equilibrium state – this can also beachieved using damped dynamics for both electrons and nuclei.
• The fictitious electronic mass m∗ must be big enough to enable the use of areasonable time step, but small enough to guarantee
– adiabaticity, i.e. no energy transfer from nuclei to electrons, which alwaysremain close to the ground state (no systematic increase of the fictitious“kinetic energy” of the electronic degrees of freedom)
– correctness of the nuclear trajectory
Typical values: m∗ ∼ 100÷ 400 electron masses
• The time step δt should be the largest value that yields a stable dynamics (nodrifts, no loss of orthonormality). Typical values: δt ∼ 0.1÷ 0.3 fs
Why (and when) Car-Parrinello works
The CP dynamics is classical, both for nuclei and electrons: the energy shouldequipartition!
Typical frequencies associated to electron dynamics: ωel ∼√
(εi − εj)/m∗.if there is a gap in the electronic spectrum, ωelmin ∼
√εgap/m∗
Typical frequencies associated to nuclear dynamics: phonon (vibrational) frequenciesωph
If ωphmax << ωelmin there is negligible energy transfer from nuclei to electrons
A fast electron dynamics keeps the electrons close to the ground state and averagesout the error on the forces. The slow nuclear dynamics is very close to the correctone, i.e. with electrons in the ground state.
(G. Pastore, E. Smargiassi, F. Buda, Phys. Rev. A 44, 6334 (1991)).
4.1 Choice of a basis set
The actual solution of the electronic problem requires to expand Kohn-Sham statesinto a sum over a suitable (finite!) set of basis functions bn(r):
ψi(r) =∑n
cnbn(r)
Typical basis sets:
• Localized functions:atom-centred functions such as
– Linear Combinations of Atomic Orbitals (LCAO)– Gaussian-type Orbitals (GTO)– Linearized Muffin-Tin Orbitals (LMTO)
• Delocalized functions:
– Plane Waves (PW)
One could also consider mixed basis sets. The Linearized Augmented Plane Waves(LAPW) could be classified in this category.
Advantages and disadvantages of various basis sets
• Localized basis sets:
+ fast convergence with respect to basis set size: few functions per atom needed+ can be used in finite as well as in periodic systems (as Bloch sums: φk =∑
R e−ik·Rφ(r−R))
– difficult to evaluate convergence quality (no systematic way to improveconvergence)
– difficult to use (two- and three-centre integrals)– difficult to calculate forces (Pulay forces if basis set is not complete)
• Plane Waves:
– slow convergence with respect to basis set size (many more PWs than localizedfunctions needed)
– require periodicity: in finite systems, supercells must be introduced+ easy to evaluate convergence quality: there is a single convergence parameter,
the cutoff)+ easy to use (Fourier transform)+ easy to calculate forces (no Pulay forces even if the basis set is incomplete)
Periodicity (or lack of it) and Supercells
Let us borrow some concepts from solid-state physics. A perfect crystal, havingdiscrete translation symmetry, is described in terms of
• a periodically repeated unit cell and a lattice oftranslation vectors R = n1R1+n2R2+n3R3, definedvia three primitive vectors R1,R2,R3 and integercoefficients n1, n2, n3.
• a basis of atomic positions di into the unit cell
• a reciprocal lattice of vectors G such thatG ·R = 2πl, with l integer:G = m1G1 +m2G2 +m3G3 with Gi ·Rj =2πδij and m1,m2,m3 integer.
For non-periodic systems, as typically found in FPMD, the periodicity is artificialand it is defined by the simulation cell, or supercell.
Band Structure, Bloch states
The one-electron states ψ(r) for a periodic system are classified by a band index iand a wave vector k:
ψi,k(r) = eik·rui,k(r)
where ui,k(r) is translationally invariant:
ui,k(r + R) = ui,k(r).
For “true” periodic system, one has to deal with sums over k-points.
In FPMD, it is usual to assume Periodic Boundary Conditions (PBC) on thesimulation cell (defined by R1,R2,R3) :
ψ(r + R1) = ψ(r + R2) = ψ(r + R3) = ψ(r).
This means that only states with k = 0 (“Gamma point”) are considered.
The simulation cell uniquely defines the PW basis set.
Plane-wave basis set
A PW basis set for states at k = 0 is defined as
〈r|G〉 =1√NΩ
eiG·r,h2
2mG2 ≤ Ecut
Ω = unit cell volume, NΩ = crystal volume, Ecut = cutoff on the kinetic energy ofPWs (in order to have a finite number of PWs!). The PW basis set is complete forEcut →∞ and orthonormal: 〈G|G′〉 = δGG′
The components on a PW basis set are the Fourier transform:
|ψi〉 =∑G
ci,G|G〉
ci,G = 〈G|ψi〉 =1√NΩ
∫ψi(r)e−i(G)·rdr = ψi(G).
Note that the larger the supercell, the larger the number of PW’s for a given cutoff.
4.2 The need for Pseudopotentials
Are PWs a practical basis set for electronic structure calculations? Not really!Simple Fourier analysis shows that Fourier components up to q ∼ 2π/δ are neededto reproduce a function that varies on a scale of length δ.
Since atoms have strongly localized core orbitals(e.g. 1s wavefunction for C has δ ' 0.1 a.u.),millions of PW’s would be needed even for simplecrystals with small cells! We need to:
• get rid of core states
• get rid of orthogonality wiggles close to thenucleus
0 0.5 1 1.5 2 2.5 3 3.5 4r (a.u.)
!1
0
1
2
R(r)
C 1s2 2s2 2p2
1s2s2p
Solution: Pseudopotentials (PP). A smooth effective potential that reproduces theeffect of the nucleus plus core electrons on valence electrons.
Understanding Pseudopotentials
Smoothness and transferability are the relevant keywords:
• We want our pseudopotential and pseudo-orbitals to be as smooth as possible sothat expansion into plane waves is convenient (i.e. the required kinetic energycutoff Ecut is small)
• We want our pseudopotential toproduce pseudo-orbitals that are asclose as possible to true (“all-electron”) orbitals outside the coreregion, for all systems containing agiven atom (in the figure: all-electronand pseudo-orbitals for Si)
Of course, the two goals are usuallyconflicting!
Pseudopotentials have a long story: let’s start from the end.
Understanding PP: Projector-Augmented Waves
Let us look for a linear operator T connecting all-electron orbitals |ψi〉 to pseudo-
orbitals |ψi〉 as in: |ψi〉 = T |ψi〉. Pseudo-orbitals will be our variational parameters.
We write the charge density, energy, etc. using pseudo-orbitals and T instead ofall-electron orbitals.
The operator T can be defined in terms of its action on atomic waves (i.e. orbitalsat a given energy, not necessarily bound states):
• |φl〉: set of atomic all-electron waves (bound or unbound states)
• |φl〉: corresponding set of atomic pseudo-waves. Beyond some suitable “core
radius” Rl, φl(r) = φl(r); for r < Rl, φl(r) are smooth functions.
(P. E. Blochl, Phys. Rev. B 50, 17953 (1994))
Understanding PP: the PAW transformation
If the above sets are complete in the core region, the operator T can be written as
|ψi〉 = T |ψi〉 = |ψi〉+∑l
(|φl〉 − |φl〉
)〈βl|ψi〉
where the βl “projectors” are atomic functions, having the properties 〈βl|φm〉 = δlmand βl(r) = 0 for r > Rl. The logic is described in the picture below:
The pseudopotential itself is written as a nonlocal operator, V , in terms of the βlprojectors:
V ≡ Vloc(r) +∑lm
|βl〉Dlm〈βm|
(Vloc contains the long-range Coulomb part −Ze2/r)
Understanding PP: Charge in PAW
The (valence) charge density is no longer the simple sum of |ψi|2:
n(r) =∑i
fi|ψi(r)|2 +∑i
fi∑lm
〈ψi|βl〉Qlm(r)〈βm|ψi〉,
andQlm(r) = φ∗l (r)φm(r)− φ∗l (r)φm(r).
The augmentation charges Qlm(r) are zero for r > Rl. A generalized orthonormalityrelation holds for pseudo-orbitals:
〈ψi|S|ψj〉 =
∫ψ∗i (r)ψj(r)dr +
∑lm
〈ψi|βl〉qlm〈βm|ψj〉 = δij
where qlm =
∫Qlm(r)dr. The Dlm quantites and βl, Qlm functions are atomic
quantities that define the PP (or PAW set).
PP taxonomy: PAW, Ultrasoft, norm-conserving
• In the full PAW scheme, the augmentation functions are calculated and stored on aradial grid, centred at each atom. The charge density is composed by a “smooth”term expanded into plane waves, and an “augmentation” term calculated on theradial grid (Kresse and Joubert, Phys. Rev. B 59, 1759 (1999))
• In the Ultrasoft PP scheme (D. Vanderbilt, B 41, R7892 (1990)), theaugmentation functions Qlm(r) are pseudized, i.e. made smoother: both“smooth” and “augmentation” terms can be calculated using FFT (see later), ineither reciprocal or real space. The augmentation term usually requires a largerFFT grid in G-space than for the smooth term (“double grid”)
• If we set Qlm(r) = 0, we obtain good old norm-conserving PPs (Hamann,Schluter, Chiang 1982) in the separable, nonlocal form.
Which pseudopotentials are good for me?
• Norm-conserving:
+ are simple to generate and to use. Theory and methodological improvementsare invariably implemented first (and often only) for norm-conserving PPs
– are relatively hard: core radii Rl cannot exceed by much the outermostmaximum of the valence atomic orbitals, or else the loss of transferability islarge. For some atoms: 2p elements C, N, O, F, 3d transition metals, 4f rareearths, this restriction may lead to very high plane-wave cutoffs (70 Ry and up)
– do not give any sensible information about the orbitals close to the nucleus(all-electron orbitals can be “reconstructed” using the PAW transformation)
Unless too hard for practical usage, this is usually the first choice
Which pseudopotentials are good for me? (II)
• Ultrasoft:
+ can be made smooth with little loss of transferability: core radii Rl can bepushed to larger values, even for “difficult cases”. Cutoffs of 25 to 35 Ry areusually good for most cases. Note that you may need a second FFT grid foraugmentation charges, with typical cutoff 8÷12× orbital cutoff (instead of 4)
- are not simple to generate: the pseudization of augmentation charges is oftena source of trouble (e.g. negative charge)
- some calculations not available for Ultrasoft PPs– give even less information about the orbitals close to the nucleus (all-electron
orbitals can be “reconstructed”)
Ultrasoft PPs are typically used in all cases wherenorm-conserving PPs are too hard: C, N, O, F, 3delements, elements with “semicore” states
Which pseudopotentials are good for me? (III)
• PAW:
+ most transferrable, even for atoms that are “difficult” for Ultrasoft PPs (e.g.magnetic materials): accuracy is comparable to all-electron techniques (e.g.FLAPW)
+ give information about the orbital close to the nucleus- less well-known and used, less experience available- not all calculations are available for PAW
Which pseudopotentials are good for me? (IV)
There are a few more aspect to be considered in the choice of a pseudopotential:
• PPs are bound to a specific XC functional, at least in principle. xception: Hybridand nonlocal (vdW-DF) functionals, for which very few (or no) PPs are available.
• The distinction between “core” and “valence” electrons is not always clear-cut.In some cases you may need to extend “valence” to include the so-called semicorestates in order to achieve better (or less lousy) transferability. E.g.: 3d states inZn and Ga; 3s and 3p states in 3d transition metals Fe, Co, Ni, ...
Inclusion of semicore states adds considerable complexity to both the generation andthe practical usage of a PP: to be done only if needed.
Where do I find pseudopotentials?
There are many ready-to-use PPs tables around, and by now several papers containingextensive PP tables, e.g.:
• PSLib (US/PAW): A. Dal Corso, Comp. Material Science 95, 337 (2014)
• PSLib tests: E. Kucukbenli et al., http://arxiv.org/abs/1404.3015
• GRBV (US/PAW): K. F. Garrity et al., Comput. Mater. Sci. 81, 446 (2014)
• PAW: F. Jollet et al., Comp. Phys. Commun. 185, 1246 (2014)
• NC: A. Willand et al., J. Chem. Phys. 138, 104109, 2013
but there is not a single standard PP file format: each code has its own format.
Pseudopotential testing
PPs must be always tested to check for— needed cutoff Ecut (plus augmentation charge cutoff for US PP)— transferability to the expected chemical environment— absence of ghost states: spurious unphysical states in the valence region ofenergies, or close to it. All nonlocal PPs can be affected
Testing can be performed
• in the single atom, using an atomic code or thePP generation code itself, by comparing energydifferences between electronic configurations, andverifying that logarithmic derivatives are wellreproduced;
• in simple molecular or solid-state systems, ideally by comparing with accurateall-electron results; less ideally, with other PP results, or with experimental data
5. Software: Quantum ESPRESSO
Car-Parrinello simulations can be performed using the CP package of quantumESPRESSO: Quantum opEn-Source Package for Research in Electronic Structure,Simulation, and Optimization.
quantum ESPRESSO is a distribution of software for atomistic calculationsbased on electronic structure, using density-functional theory, a plane-wave basisset, pseudopotentials. Freely available under the terms of the GNU General PublicLicense.
Main goals of quantum ESPRESSO are innovation in methods and algorithms;efficiency on modern computer architectures.
quantum ESPRESSO implements multiple parallelization levels
See: http://www.quantum-espresso.org for more info, in particularhttp://www.quantum-espresso.org/pseudopotentials/ for available PP’s.
www.quantum-espresso.org
CP package
Car-Parrinello variable-cell molecular dynamics for norm-conserving or ultrasoft PPs,Γ point (k = 0) only. Originaly developed by A. Pasquarello (EPF Lausanne), K.Laasonen (Oulu), A. Trave (LLNL), R. Car (Princeton), P. Giannozzi (Udine), N.Marzari (EPF Lausanne); C. Cavazzoni (CINECA), S. Scandolo (ICTP), and manyothers.
• Electronic and ionic minimization schemes: damped dynamics, conjugate gradient
• Verlet dynamics with mass preconditioning
• Constrained dynamics
• Nose thermostat for both electrons and nuclei, velocity rescaling
• Fast treatment (“grid box”) of augmentation terms in Ultrasoft PPs
• Parallelization levels: on plane-wave and real-space grids; on Kohn-Sham states;on FFT’s in Hψi products (“task groups”); OpenMP
CP package, advanced features
• Nonlocal van-der-Waals functionals, or semiempirical corrections
• DFT+U and metaGGA functionals
• Fast implementation of hybrid functionals with Wannier functions
• Modified kinetic functional for constant-pressure calculations
• Metallic systems: variable-occupancy (ensemble) dynamics
• Self-Interaction Correction for systems with one unpaired electron
• Dynamics with Wannier functions under an external electric field
• Finite electric fields with Berry’s phase
• With plumed plugin: metadynamics (Laio-Parrinello)
Some words on computer requirements
Quantum simulations are both CPU and RAM-intensive.Actual CPU time and RAM requirements depend upon:
• size of the system under examination: As a rule of thumb, CPU ∝ N2÷3, RAM∝ N2, where N = number of atoms
• kind of system: type and arrangement of atoms, influencing the number of planewaves, of electronic states, of k-points needed...
• desired results: computational effort increases from simple self-consistent (single-point) calculation to structural optimization to reaction pathways, molecular-dynamics simulations, ...
CPU time mostly spent in FFT and linear algebra.RAM mostly needed to store Kohn-Sham orbitals.
Typical computational requirements
For typical biological systems (O(1000) atoms): days to weeks of CPU, or more,tens to hundreds Gb RAM. Massively parallel machines are needed, together witheffective, memory-distributing algorithms.
Fragment of an Aβ-peptide in water containing 838 atoms and 2312 electronsin a 22.1×22.9×19.9 A3 cell, Γ-point. CP code on BlueGene/P, 4 processes percomputing node.
Very Technical Appendix:Fast Fourier Transforms, Dual-space technique
Let us consider first the simple case of a periodical function f(x), with period L. ItsFourier components:
f(q) =1
L
∫f(x)e−iqxdx
are nonzero over an infinite set of discrete values of q:
qn = n2π
L, −∞ < n <∞
The Fourier components decay to 0 for large q. The inverse Fourier transform hasthe form
f(x) =∑n
f(qn)eiqnx =∑n
fnein(2πx/L)
Our functions are however defined over a discrete but finite grid of q. How are theyrepresented in x space?
Discrete Fourier Transform
We assume both x and q grids to be discrete, finite, and periodically repeated, andwe write, for N large enough to accommodate all q components:
f(x) → fm = f(xm), xm = mL
N, m = 0, .., N − 1
f(q) → fn = f(qn), qn = n2π
L, n = 0, .., N − 1
(q components of negative value refold into those at the end of the box). TheDiscrete Fourier Transform can be written as
fm =
N−1∑n=0
fnei(2πnm/N) (x− space)
fn =1
N
N−1∑m=0
fme−i(2πnm/N) (q − space)
Discrete Fourier Transform in 3D
Generalization of the Discrete Fourier Transform to 3 dimensions:
G = n′1G1 + n′2G2 + n′3G3
with n1 = 0, .., N1 − 1, n2 = 0, .., N2 − 1, n3 = 0, .., N3 − 1, and n′1, n′2, n′3 are
n1, n2, n3 refolded so that they are centered around the origin (remember: theG−space grid is also periodic!). G1,G2,G3 are the primitive translations of theunit cell of the reciprocal lattice.
r = m1R1
N1+m2
R2
N2+m3
R3
N3
with m1 = 0, .., N1 − 1, m2 = 0, .., N2 − 1, m3 = 0, .., N3 − 1. R1,R2,R3 are theprimitive translations of the unit cell. This grid spans the unit cell.
N1, N2, N3 define the FFT grid.
Discrete Fourier Transform in 3D (2)
Original Fourier transform:
f(r) =∑G
f(G)eiG·r → f(m1,m2,m3)
f(G) =1
Ω
∫f(r)e−iG·rdr→ f(n1, n2, n3)
Discretized Fourier Transform:
f(m1,m2,m3) =∑
n1,n2,n3
f(n1, n2, n3)ei(2πn1m1/N1)ei(2πn2m2/N2)ei(2πn3m3/N3)
f(n1, n2, n3) =1
N
∑m1,m2,m3
f(m1,m2,m3)e−i(2πn1m1/N1)e−i(2πn2m2/N2)e−i(2πn3m3/N3)
where N = N1N2N3. Remember that Gi ·Rj = 2πδij.
PW-PP calculations and Discrete Fourier Transform
ψi(r) =1
V
∑G
ck+Gei(k+G)·r,
h2
2m|k + G|2 ≤ Ecut
Which grid in G-space? Need to calculate the charge density. From its G-spaceexpression:
n(G′) =∑G
∑i,k
fi,kc∗i,k+Gci,k+G+G′
one can see that Fourier components G′ up to max(|G′|) = 2max(|G|) appear. Orwe need the product of the potential time a wavefunction:
(V ψ)(G) =∑G′
V (G−G′)ci,k+G′
Again, max(|G−G′|) = 2max(|G|). We need a kinetic energy cutoff for the Fouriercomponents of the charge and potentials that is four time larger as the cutoff forthe PW basis set:
h2
2m|G|2 ≤ 4Ecut
In practice such condition may occasionally be relaxed. Important: for ultrasoftpseudopotentials, a different (larger) cutoff for augmentation charges may be needed!
The Fourier Transform grid is thus
G = n′1G1 + n′2G2 + n′3G3
with n1 = 0, .., N1 − 1, n2 = 0, .., N2 − 1, n3 = 0, .., N3 − 1. This grid must be bigenough to include all G−vectors up to a cutoff
h2
2m|G|2 ≤ 4Ecut
and NOT up to the cutoff of the PW basis set! In general, the grid will alsocontain “useless” Fourier components (beyond the above-mentioned cutoff, so thatn(G) = 0, V (G) = 0 etc.)
Fast Fourier Transform (FFT): allows to perform a Discrete Fourier Transform oforder n with a computational cost TCPU = O(n log n) (instead of O(n2)).
Advantages of the use of FFT in PW-PP calculations: enormous, especially inconjunction with iterative techniques and of the “dual-space” technique
FFT grid
(Note: G2/2 is the kinetic energy in Hartree atomic units)
Dual space technique
The most important ingredient of a PW-PP calculations is
Hψ ≡ (T + VNL + Vloc + VH + Vxc)ψ
(Tψ) : easy in G-space, TCPU = O(N)
(Vloc + VH + Vxc)ψ : easy in r-space, TCPU = O(N)
(VNLψ) : easy in G-space (also in r-space) if V is written in separable formTCPU = O(mN), m =number of projectors
FFT is used to jump from real to reciprocal space.Operations are performed in the space where it is more convenient.
The same technique is used to calculate the charge density from Kohn-Sham orbitals,the exchange-correlation GGA potential from the charge density, etc.: in all cases,we move to the more convenient space to perform the required operations.